Agentic Harnesses in High-Stakes Workflows: Inside Cooley's GO Public and Hex's GPT-6 Astra Integration
Updated: Sep 21
Professional services and data platforms are moving beyond simple chatbots to structured agentic systems that automate complex knowledge work while enforcing human oversight.
Key Findings
1. Cooley has developed GO Public, a proprietary agentic system built on ChatGPT Work, to accelerate initial public offering (IPO) preparation by shifting from generic precedents to client-specific data synthesis [1].
2. The GO Public system utilizes an agentic harness to enforce a controlled workflow, defining automated steps while mandating human lawyer validation and review [1].
3. Startup data platform Hex is integrating GPT-6 Astra to power data agents that perform complex geospatial and trend visualizations while exercising analytical judgment to verify if answers align with business objectives [2].
4. These deployments signal a structural shift in professional services, moving from simple prompt-and-response interactions to structured, multi-agent architectures designed to manage high-liability, information-dense tasks [1] [2].
The integration of artificial intelligence into high-stakes corporate environments is undergoing a structural transition. Rather than relying on generic conversational interfaces, enterprise early adopters are building structured, multi-agent architectures designed to handle complex, information-dense workflows.
Two prominent examples of this shift emerged in September 2026: international law firm Cooley announced its GO Public system built on OpenAI's ChatGPT Work [1], while data platform Hex revealed its integration of GPT-6 Astra to drive automated data visualization and analysis [2].
Both implementations demonstrate how professional services are moving toward agentic harnesses—software frameworks that constrain, guide, and validate AI agents to ensure their outputs meet rigorous professional standards [1] [2].
Historically, the adoption of generative AI in professional services was limited by the technology's tendency to hallucinate, its lack of domain-specific context, and the absence of structured workflows. In fields like capital markets law or enterprise data analytics, errors carry severe financial and regulatory consequences.
Consequently, early enterprise use cases were often restricted to low-risk tasks such as drafting basic emails or summarizing internal memos.
The development of specialized agentic systems represents a departure from this prior state, introducing controlled environments where AI agents execute multi-step processes under strict human supervision [1] [2].
The Architecture of Cooley's GO Public
Cooley, an international law firm that advised on 180 global deals totaling more than $51.5 billion in volume in 2025, developed GO Public to address the labor-intensive nature of initial public offering (IPO) preparation [1].
Typically, preparing a company for an IPO involves synthesizing thousands of disparate data points, reviewing historical disclosures, and drafting extensive registration statements [1].
According to David Wang, Chief Innovation Officer at Cooley, the sheer volume of information requires constant, intensive manual effort from legal teams [1].
To optimize this process, Cooley's legal engineers, innovation counsel, and practitioners built a proprietary agentic harness on top of ChatGPT Work [1]. This harness acts as a controlled workflow engine. Instead of allowing the underlying AI model to operate without boundaries, the harness defines specific, sequential steps that the agents can perform automatically [1].
Crucially, it establishes explicit checkpoints where human lawyers must review, validate, and sign off on the work before the process can proceed [1]. This design ensures that the firm's professional judgment and legal accountability remain central to the transaction [1].
The system also changes how legal drafting begins. In the traditional workflow, legal teams typically started by identifying a precedent transaction from a comparable company and manually adapting it to fit the new client's circumstances [1].
GO Public reverses this approach. The system begins with the client's own provided information, combines it with relevant public sources, and integrates carefully curated precedents containing what Wang describes as baked-in know-how [1].
The result is a highly tailored, client-specific starting point that lawyers can immediately begin refining, reducing the time spent on initial drafting and manual information sorting [1].
Hex and GPT-6 Astra: Engineering Analytical Judgment
While Cooley applies agentic workflows to legal text, startup data platform Hex is utilizing OpenAI's GPT-6 Astra to solve long-standing challenges in data visualization and analysis [2]. According to Caitlin Colgrove, co-founder and CTO of Hex, AI models have historically struggled to generate accurate and aesthetically pleasing data visualizations [2].
Producing a high-quality visual report requires not only writing correct code but also making complex design choices and executing intricate data transformations, such as those required for geospatial mapping [2].
Hex leverages GPT-6 Astra to power data agents that can navigate underlying code libraries, perform complex data transformations, and generate interactive dashboards [2]. For example, a user can query the system about sales channel performance over time [2].
The agentic system then generates written findings, trend lines, comparative metrics, and interactive geographic breakdowns [2]. This allows non-technical business users to interact with and explore complex datasets without needing to write SQL or Python code directly [2].
A critical component of the Hex implementation is the concept of analytical judgment [2].
Colgrove notes that even syntactically correct code can produce visualizations or answers that miss the broader business context or fail to address the user's actual question [2].
To mitigate this, Hex uses GPT-6 Astra to interrogate its own outputs [2]. The model evaluates whether the generated numbers make sense, whether the analysis directly answers the user's prompt, and whether the output aligns with the stated business objective [2]. This self-interrogation layer represents an important step toward autonomous quality control in AI-driven analytics [2].
Strategic Implications and Professional Liability
The deployment of these systems highlights a broader shift in how professional expertise is leveraged. Dave Peinsipp, partner and co-chair of Cooley's global capital markets group, emphasizes that the goal of GO Public is speed to quality rather than simple automation [1].
By automating the initial synthesis of information, legal teams can redirect their time toward high-value tasks that require strategic thinking, market experience, and professional judgment—such as challenging disclosures and advising management teams on critical transaction decisions [1]. Similarly, for corporate executives, faster IPO preparation reduces the administrative burden, allowing them to focus on running their businesses during the intensive transition to public markets [1].
However, the integration of agentic systems into high-stakes environments also introduces distinct operational risks and unanswered questions. In legal and financial compliance, the margin for error is virtually zero. While Cooley's agentic harness is designed to enforce human validation at every step, the risk of automation bias—where human reviewers become overly trusting of AI-generated drafts—remains a critical concern for risk management officers. If a lawyer overlooks an error introduced by the system, the firm still bears full professional liability for the resulting filing [1].
Furthermore, the reliance on proprietary, closed-source models like GPT-6 Astra and ChatGPT Work introduces platform dependencies [1] [2]. Organizations must weigh the efficiency gains of these tools against potential concerns regarding data privacy, long-term API pricing stability, and the risk of model drift, where updates to the underlying AI model alter its behavior and disrupt established agentic workflows.
What to Watch Next
As these technologies mature, market observers should monitor several key indicators of adoption and performance. First, the expansion of agentic frameworks beyond initial public offerings into broader capital markets transactions, such as debt offerings and mergers and acquisitions, will indicate the scalability of Cooley's GO Public model [1].
Second, the rate at which competing professional services firms develop their own proprietary harnesses will determine whether AI-driven preparation becomes a standard industry expectation or remains a competitive differentiator.
Additionally, the performance of GPT-6 Astra in diverse enterprise environments will clarify whether analytical judgment can be reliably maintained across different industries and data structures [2]. If data agents can consistently validate their own work and produce reliable, interactive reports, the barrier to sophisticated data analysis will continue to fall, accelerating the democratization of data across corporate enterprises [2].
Sources
[1] OpenAI — How Cooley is accelerating IPO work with ChatGPT
2026-09-18
2026-09-14
SteelGate note: Current-event claims in this article are grounded in the verified first-party sources listed above. Background explanation is provided for context and should be distinguished from source-attributed claims.
Comments